mm/khugepaged.c

   1 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   2
   3 #include <linux/mm.h>
   4 #include <linux/sched.h>
   5 #include <linux/mmu_notifier.h>
   6 #include <linux/rmap.h>
   7 #include <linux/swap.h>
   8 #include <linux/mm_inline.h>
   9 #include <linux/kthread.h>
  10 #include <linux/khugepaged.h>
  11 #include <linux/freezer.h>
  12 #include <linux/mman.h>
  13 #include <linux/hashtable.h>
  14 #include <linux/userfaultfd_k.h>
  15 #include <linux/page_idle.h>
  16 #include <linux/swapops.h>
  17 #include <linux/shmem_fs.h>
  18
  19 #include <asm/tlb.h>
  20 #include <asm/pgalloc.h>
  21 #include "internal.h"
  22
  23 enum scan_result {
  24         SCAN_FAIL,
  25         SCAN_SUCCEED,
  26         SCAN_PMD_NULL,
  27         SCAN_EXCEED_NONE_PTE,
  28         SCAN_PTE_NON_PRESENT,
  29         SCAN_PAGE_RO,
  30         SCAN_LACK_REFERENCED_PAGE,
  31         SCAN_PAGE_NULL,
  32         SCAN_SCAN_ABORT,
  33         SCAN_PAGE_COUNT,
  34         SCAN_PAGE_LRU,
  35         SCAN_PAGE_LOCK,
  36         SCAN_PAGE_ANON,
  37         SCAN_PAGE_COMPOUND,
  38         SCAN_ANY_PROCESS,
  39         SCAN_VMA_NULL,
  40         SCAN_VMA_CHECK,
  41         SCAN_ADDRESS_RANGE,
  42         SCAN_SWAP_CACHE_PAGE,
  43         SCAN_DEL_PAGE_LRU,
  44         SCAN_ALLOC_HUGE_PAGE_FAIL,
  45         SCAN_CGROUP_CHARGE_FAIL,
  46         SCAN_EXCEED_SWAP_PTE,
  47         SCAN_TRUNCATED,
  48 };
  49
  50 #define CREATE_TRACE_POINTS
  51 #include <trace/events/huge_memory.h>
  52
  53 static struct task_struct *khugepaged_thread __read_mostly;
  54 static DEFINE_MUTEX(khugepaged_mutex);
  55
  56 /* default scan 8*512 pte (or vmas) every 30 second */
  57 static unsigned int khugepaged_pages_to_scan __read_mostly;
  58 static unsigned int khugepaged_pages_collapsed;
  59 static unsigned int khugepaged_full_scans;
  60 static unsigned int khugepaged_scan_sleep_millisecs __read_mostly = 10000;
  61 /* during fragmentation poll the hugepage allocator once every minute */
  62 static unsigned int khugepaged_alloc_sleep_millisecs __read_mostly = 60000;
  63 static unsigned long khugepaged_sleep_expire;
  64 static DEFINE_SPINLOCK(khugepaged_mm_lock);
  65 static DECLARE_WAIT_QUEUE_HEAD(khugepaged_wait);
  66 /*
  67  * default collapse hugepages if there is at least one pte mapped like
  68  * it would have happened if the vma was large enough during page
  69  * fault.
  70  */
  71 static unsigned int khugepaged_max_ptes_none __read_mostly;
  72 static unsigned int khugepaged_max_ptes_swap __read_mostly;
  73
  74 #define MM_SLOTS_HASH_BITS 10
  75 static __read_mostly DEFINE_HASHTABLE(mm_slots_hash, MM_SLOTS_HASH_BITS);
  76
  77 static struct kmem_cache *mm_slot_cache __read_mostly;
  78
  79 /**
  80  * struct mm_slot - hash lookup from mm to mm_slot
  81  * @hash: hash collision list
  82  * @mm_node: khugepaged scan list headed in khugepaged_scan.mm_head
  83  * @mm: the mm that this information is valid for
  84  */
  85 struct mm_slot {
  86         struct hlist_node hash;
  87         struct list_head mm_node;
  88         struct mm_struct *mm;
  89 };
  90
  91 /**
  92  * struct khugepaged_scan - cursor for scanning
  93  * @mm_head: the head of the mm list to scan
  94  * @mm_slot: the current mm_slot we are scanning
  95  * @address: the next address inside that to be scanned
  96  *
  97  * There is only the one khugepaged_scan instance of this cursor structure.
  98  */
  99 struct khugepaged_scan {
 100         struct list_head mm_head;
 101         struct mm_slot *mm_slot;
 102         unsigned long address;
 103 };
 104
 105 static struct khugepaged_scan khugepaged_scan = {
 106         .mm_head = LIST_HEAD_INIT(khugepaged_scan.mm_head),
 107 };
 108
 109 #ifdef CONFIG_SYSFS
 110 static ssize_t scan_sleep_millisecs_show(struct kobject *kobj,
 111                                          struct kobj_attribute *attr,
 112                                          char *buf)
 113 {
 114         return sprintf(buf, "%u\n", khugepaged_scan_sleep_millisecs);
 115 }
 116
 117 static ssize_t scan_sleep_millisecs_store(struct kobject *kobj,
 118                                           struct kobj_attribute *attr,
 119                                           const char *buf, size_t count)
 120 {
 121         unsigned long msecs;
 122         int err;
 123
 124         err = kstrtoul(buf, 10, &msecs);
 125         if (err || msecs > UINT_MAX)
 126                 return -EINVAL;
 127
 128         khugepaged_scan_sleep_millisecs = msecs;
 129         khugepaged_sleep_expire = 0;
 130         wake_up_interruptible(&khugepaged_wait);
 131
 132         return count;
 133 }
 134 static struct kobj_attribute scan_sleep_millisecs_attr =
 135         __ATTR(scan_sleep_millisecs, 0644, scan_sleep_millisecs_show,
 136                scan_sleep_millisecs_store);
 137
 138 static ssize_t alloc_sleep_millisecs_show(struct kobject *kobj,
 139                                           struct kobj_attribute *attr,
 140                                           char *buf)
 141 {
 142         return sprintf(buf, "%u\n", khugepaged_alloc_sleep_millisecs);
 143 }
 144
 145 static ssize_t alloc_sleep_millisecs_store(struct kobject *kobj,
 146                                            struct kobj_attribute *attr,
 147                                            const char *buf, size_t count)
 148 {
 149         unsigned long msecs;
 150         int err;
 151
 152         err = kstrtoul(buf, 10, &msecs);
 153         if (err || msecs > UINT_MAX)
 154                 return -EINVAL;
 155
 156         khugepaged_alloc_sleep_millisecs = msecs;
 157         khugepaged_sleep_expire = 0;
 158         wake_up_interruptible(&khugepaged_wait);
 159
 160         return count;
 161 }
 162 static struct kobj_attribute alloc_sleep_millisecs_attr =
 163         __ATTR(alloc_sleep_millisecs, 0644, alloc_sleep_millisecs_show,
 164                alloc_sleep_millisecs_store);
 165
 166 static ssize_t pages_to_scan_show(struct kobject *kobj,
 167                                   struct kobj_attribute *attr,
 168                                   char *buf)
 169 {
 170         return sprintf(buf, "%u\n", khugepaged_pages_to_scan);
 171 }
 172 static ssize_t pages_to_scan_store(struct kobject *kobj,
 173                                    struct kobj_attribute *attr,
 174                                    const char *buf, size_t count)
 175 {
 176         int err;
 177         unsigned long pages;
 178
 179         err = kstrtoul(buf, 10, &pages);
 180         if (err || !pages || pages > UINT_MAX)
 181                 return -EINVAL;
 182
 183         khugepaged_pages_to_scan = pages;
 184
 185         return count;
 186 }
 187 static struct kobj_attribute pages_to_scan_attr =
 188         __ATTR(pages_to_scan, 0644, pages_to_scan_show,
 189                pages_to_scan_store);
 190
 191 static ssize_t pages_collapsed_show(struct kobject *kobj,
 192                                     struct kobj_attribute *attr,
 193                                     char *buf)
 194 {
 195         return sprintf(buf, "%u\n", khugepaged_pages_collapsed);
 196 }
 197 static struct kobj_attribute pages_collapsed_attr =
 198         __ATTR_RO(pages_collapsed);
 199
 200 static ssize_t full_scans_show(struct kobject *kobj,
 201                                struct kobj_attribute *attr,
 202                                char *buf)
 203 {
 204         return sprintf(buf, "%u\n", khugepaged_full_scans);
 205 }
 206 static struct kobj_attribute full_scans_attr =
 207         __ATTR_RO(full_scans);
 208
 209 static ssize_t khugepaged_defrag_show(struct kobject *kobj,
 210                                       struct kobj_attribute *attr, char *buf)
 211 {
 212         return single_hugepage_flag_show(kobj, attr, buf,
 213                                 TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 214 }
 215 static ssize_t khugepaged_defrag_store(struct kobject *kobj,
 216                                        struct kobj_attribute *attr,
 217                                        const char *buf, size_t count)
 218 {
 219         return single_hugepage_flag_store(kobj, attr, buf, count,
 220                                  TRANSPARENT_HUGEPAGE_DEFRAG_KHUGEPAGED_FLAG);
 221 }
 222 static struct kobj_attribute khugepaged_defrag_attr =
 223         __ATTR(defrag, 0644, khugepaged_defrag_show,
 224                khugepaged_defrag_store);
 225
 226 /*
 227  * max_ptes_none controls if khugepaged should collapse hugepages over
 228  * any unmapped ptes in turn potentially increasing the memory
 229  * footprint of the vmas. When max_ptes_none is 0 khugepaged will not
 230  * reduce the available free memory in the system as it
 231  * runs. Increasing max_ptes_none will instead potentially reduce the
 232  * free memory in the system during the khugepaged scan.
 233  */
 234 static ssize_t khugepaged_max_ptes_none_show(struct kobject *kobj,
 235                                              struct kobj_attribute *attr,
 236                                              char *buf)
 237 {
 238         return sprintf(buf, "%u\n", khugepaged_max_ptes_none);
 239 }
 240 static ssize_t khugepaged_max_ptes_none_store(struct kobject *kobj,
 241                                               struct kobj_attribute *attr,
 242                                               const char *buf, size_t count)
 243 {
 244         int err;
 245         unsigned long max_ptes_none;
 246
 247         err = kstrtoul(buf, 10, &max_ptes_none);
 248         if (err || max_ptes_none > HPAGE_PMD_NR-1)
 249                 return -EINVAL;
 250
 251         khugepaged_max_ptes_none = max_ptes_none;
 252
 253         return count;
 254 }
 255 static struct kobj_attribute khugepaged_max_ptes_none_attr =
 256         __ATTR(max_ptes_none, 0644, khugepaged_max_ptes_none_show,
 257                khugepaged_max_ptes_none_store);
 258
 259 static ssize_t khugepaged_max_ptes_swap_show(struct kobject *kobj,
 260                                              struct kobj_attribute *attr,
 261                                              char *buf)
 262 {
 263         return sprintf(buf, "%u\n", khugepaged_max_ptes_swap);
 264 }
 265
 266 static ssize_t khugepaged_max_ptes_swap_store(struct kobject *kobj,
 267                                               struct kobj_attribute *attr,
 268                                               const char *buf, size_t count)
 269 {
 270         int err;
 271         unsigned long max_ptes_swap;
 272
 273         err  = kstrtoul(buf, 10, &max_ptes_swap);
 274         if (err || max_ptes_swap > HPAGE_PMD_NR-1)
 275                 return -EINVAL;
 276
 277         khugepaged_max_ptes_swap = max_ptes_swap;
 278
 279         return count;
 280 }
 281
 282 static struct kobj_attribute khugepaged_max_ptes_swap_attr =
 283         __ATTR(max_ptes_swap, 0644, khugepaged_max_ptes_swap_show,
 284                khugepaged_max_ptes_swap_store);
 285
 286 static struct attribute *khugepaged_attr[] = {
 287         &khugepaged_defrag_attr.attr,
 288         &khugepaged_max_ptes_none_attr.attr,
 289         &pages_to_scan_attr.attr,
 290         &pages_collapsed_attr.attr,
 291         &full_scans_attr.attr,
 292         &scan_sleep_millisecs_attr.attr,
 293         &alloc_sleep_millisecs_attr.attr,
 294         &khugepaged_max_ptes_swap_attr.attr,
 295         NULL,
 296 };
 297
 298 struct attribute_group khugepaged_attr_group = {
 299         .attrs = khugepaged_attr,
 300         .name = "khugepaged",
 301 };
 302 #endif /* CONFIG_SYSFS */
 303
 304 #define VM_NO_KHUGEPAGED (VM_SPECIAL | VM_HUGETLB)
 305
 306 int hugepage_madvise(struct vm_area_struct *vma,
 307                      unsigned long *vm_flags, int advice)
 308 {
 309         switch (advice) {
 310         case MADV_HUGEPAGE:
 311 #ifdef CONFIG_S390
 312                 /*
 313                  * qemu blindly sets MADV_HUGEPAGE on all allocations, but s390
 314                  * can't handle this properly after s390_enable_sie, so we simply
 315                  * ignore the madvise to prevent qemu from causing a SIGSEGV.
 316                  */
 317                 if (mm_has_pgste(vma->vm_mm))
 318                         return 0;
 319 #endif
 320                 *vm_flags &= ~VM_NOHUGEPAGE;
 321                 *vm_flags |= VM_HUGEPAGE;
 322                 /*
 323                  * If the vma become good for khugepaged to scan,
 324                  * register it here without waiting a page fault that
 325                  * may not happen any time soon.
 326                  */
 327                 if (!(*vm_flags & VM_NO_KHUGEPAGED) &&
 328                                 khugepaged_enter_vma_merge(vma, *vm_flags))
 329                         return -ENOMEM;
 330                 break;
 331         case MADV_NOHUGEPAGE:
 332                 *vm_flags &= ~VM_HUGEPAGE;
 333                 *vm_flags |= VM_NOHUGEPAGE;
 334                 /*
 335                  * Setting VM_NOHUGEPAGE will prevent khugepaged from scanning
 336                  * this vma even if we leave the mm registered in khugepaged if
 337                  * it got registered before VM_NOHUGEPAGE was set.
 338                  */
 339                 break;
 340         }
 341
 342         return 0;
 343 }
 344
 345 int __init khugepaged_init(void)
 346 {
 347         mm_slot_cache = kmem_cache_create("khugepaged_mm_slot",
 348                                           sizeof(struct mm_slot),
 349                                           __alignof__(struct mm_slot), 0, NULL);
 350         if (!mm_slot_cache)
 351                 return -ENOMEM;
 352
 353         khugepaged_pages_to_scan = HPAGE_PMD_NR * 8;
 354         khugepaged_max_ptes_none = HPAGE_PMD_NR - 1;
 355         khugepaged_max_ptes_swap = HPAGE_PMD_NR / 8;
 356
 357         return 0;
 358 }
 359
 360 void __init khugepaged_destroy(void)
 361 {
 362         kmem_cache_destroy(mm_slot_cache);
 363 }
 364
 365 static inline struct mm_slot *alloc_mm_slot(void)
 366 {
 367         if (!mm_slot_cache)     /* initialization failed */
 368                 return NULL;
 369         return kmem_cache_zalloc(mm_slot_cache, GFP_KERNEL);
 370 }
 371
 372 static inline void free_mm_slot(struct mm_slot *mm_slot)
 373 {
 374         kmem_cache_free(mm_slot_cache, mm_slot);
 375 }
 376
 377 static struct mm_slot *get_mm_slot(struct mm_struct *mm)
 378 {
 379         struct mm_slot *mm_slot;
 380
 381         hash_for_each_possible(mm_slots_hash, mm_slot, hash, (unsigned long)mm)
 382                 if (mm == mm_slot->mm)
 383                         return mm_slot;
 384
 385         return NULL;
 386 }
 387
 388 static void insert_to_mm_slots_hash(struct mm_struct *mm,
 389                                     struct mm_slot *mm_slot)
 390 {
 391         mm_slot->mm = mm;
 392         hash_add(mm_slots_hash, &mm_slot->hash, (long)mm);
 393 }
 394
 395 static inline int khugepaged_test_exit(struct mm_struct *mm)
 396 {
 397         return atomic_read(&mm->mm_users) == 0 || !mmget_still_valid(mm);
 398 }
 399
 400 int __khugepaged_enter(struct mm_struct *mm)
 401 {
 402         struct mm_slot *mm_slot;
 403         int wakeup;
 404
 405         mm_slot = alloc_mm_slot();
 406         if (!mm_slot)
 407                 return -ENOMEM;
 408
 409         /* __khugepaged_exit() must not run from under us */
 410         VM_BUG_ON_MM(atomic_read(&mm->mm_users) == 0, mm);
 411         if (unlikely(test_and_set_bit(MMF_VM_HUGEPAGE, &mm->flags))) {
 412                 free_mm_slot(mm_slot);
 413                 return 0;
 414         }
 415
 416         spin_lock(&khugepaged_mm_lock);
 417         insert_to_mm_slots_hash(mm, mm_slot);
 418         /*
 419          * Insert just behind the scanning cursor, to let the area settle
 420          * down a little.
 421          */
 422         wakeup = list_empty(&khugepaged_scan.mm_head);
 423         list_add_tail(&mm_slot->mm_node, &khugepaged_scan.mm_head);
 424         spin_unlock(&khugepaged_mm_lock);
 425
 426         atomic_inc(&mm->mm_count);
 427         if (wakeup)
 428                 wake_up_interruptible(&khugepaged_wait);
 429
 430         return 0;
 431 }
 432
 433 int khugepaged_enter_vma_merge(struct vm_area_struct *vma,
 434                                unsigned long vm_flags)
 435 {
 436         unsigned long hstart, hend;
 437         if (!vma->anon_vma)
 438                 /*
 439                  * Not yet faulted in so we will register later in the
 440                  * page fault if needed.
 441                  */
 442                 return 0;
 443         if (vma->vm_ops || (vm_flags & VM_NO_KHUGEPAGED))
 444                 /* khugepaged not yet working on file or special mappings */
 445                 return 0;
 446         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 447         hend = vma->vm_end & HPAGE_PMD_MASK;
 448         if (hstart < hend)
 449                 return khugepaged_enter(vma, vm_flags);
 450         return 0;
 451 }
 452
 453 void __khugepaged_exit(struct mm_struct *mm)
 454 {
 455         struct mm_slot *mm_slot;
 456         int free = 0;
 457
 458         spin_lock(&khugepaged_mm_lock);
 459         mm_slot = get_mm_slot(mm);
 460         if (mm_slot && khugepaged_scan.mm_slot != mm_slot) {
 461                 hash_del(&mm_slot->hash);
 462                 list_del(&mm_slot->mm_node);
 463                 free = 1;
 464         }
 465         spin_unlock(&khugepaged_mm_lock);
 466
 467         if (free) {
 468                 clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
 469                 free_mm_slot(mm_slot);
 470                 mmdrop(mm);
 471         } else if (mm_slot) {
 472                 /*
 473                  * This is required to serialize against
 474                  * khugepaged_test_exit() (which is guaranteed to run
 475                  * under mmap sem read mode). Stop here (after we
 476                  * return all pagetables will be destroyed) until
 477                  * khugepaged has finished working on the pagetables
 478                  * under the mmap_sem.
 479                  */
 480                 down_write(&mm->mmap_sem);
 481                 up_write(&mm->mmap_sem);
 482         }
 483 }
 484
 485 static void release_pte_page(struct page *page)
 486 {
 487         /* 0 stands for page_is_file_cache(page) == false */
 488         dec_node_page_state(page, NR_ISOLATED_ANON + 0);
 489         unlock_page(page);
 490         putback_lru_page(page);
 491 }
 492
 493 static void release_pte_pages(pte_t *pte, pte_t *_pte)
 494 {
 495         while (--_pte >= pte) {
 496                 pte_t pteval = *_pte;
 497                 if (!pte_none(pteval) && !is_zero_pfn(pte_pfn(pteval)))
 498                         release_pte_page(pte_page(pteval));
 499         }
 500 }
 501
 502 static int __collapse_huge_page_isolate(struct vm_area_struct *vma,
 503                                         unsigned long address,
 504                                         pte_t *pte)
 505 {
 506         struct page *page = NULL;
 507         pte_t *_pte;
 508         int none_or_zero = 0, result = 0, referenced = 0;
 509         bool writable = false;
 510
 511         for (_pte = pte; _pte < pte+HPAGE_PMD_NR;
 512              _pte++, address += PAGE_SIZE) {
 513                 pte_t pteval = *_pte;
 514                 if (pte_none(pteval) || (pte_present(pteval) &&
 515                                 is_zero_pfn(pte_pfn(pteval)))) {
 516                         if (!userfaultfd_armed(vma) &&
 517                             ++none_or_zero <= khugepaged_max_ptes_none) {
 518                                 continue;
 519                         } else {
 520                                 result = SCAN_EXCEED_NONE_PTE;
 521                                 goto out;
 522                         }
 523                 }
 524                 if (!pte_present(pteval)) {
 525                         result = SCAN_PTE_NON_PRESENT;
 526                         goto out;
 527                 }
 528                 page = vm_normal_page(vma, address, pteval);
 529                 if (unlikely(!page)) {
 530                         result = SCAN_PAGE_NULL;
 531                         goto out;
 532                 }
 533
 534                 /* TODO: teach khugepaged to collapse THP mapped with pte */
 535                 if (PageCompound(page)) {
 536                         result = SCAN_PAGE_COMPOUND;
 537                         goto out;
 538                 }
 539
 540                 VM_BUG_ON_PAGE(!PageAnon(page), page);
 541                 VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
 542
 543                 /*
 544                  * We can do it before isolate_lru_page because the
 545                  * page can't be freed from under us. NOTE: PG_lock
 546                  * is needed to serialize against split_huge_page
 547                  * when invoked from the VM.
 548                  */
 549                 if (!trylock_page(page)) {
 550                         result = SCAN_PAGE_LOCK;
 551                         goto out;
 552                 }
 553
 554                 /*
 555                  * cannot use mapcount: can't collapse if there's a gup pin.
 556                  * The page must only be referenced by the scanned process
 557                  * and page swap cache.
 558                  */
 559                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
 560                         unlock_page(page);
 561                         result = SCAN_PAGE_COUNT;
 562                         goto out;
 563                 }
 564                 if (pte_write(pteval)) {
 565                         writable = true;
 566                 } else {
 567                         if (PageSwapCache(page) &&
 568                             !reuse_swap_page(page, NULL)) {
 569                                 unlock_page(page);
 570                                 result = SCAN_SWAP_CACHE_PAGE;
 571                                 goto out;
 572                         }
 573                         /*
 574                          * Page is not in the swap cache. It can be collapsed
 575                          * into a THP.
 576                          */
 577                 }
 578
 579                 /*
 580                  * Isolate the page to avoid collapsing an hugepage
 581                  * currently in use by the VM.
 582                  */
 583                 if (isolate_lru_page(page)) {
 584                         unlock_page(page);
 585                         result = SCAN_DEL_PAGE_LRU;
 586                         goto out;
 587                 }
 588                 /* 0 stands for page_is_file_cache(page) == false */
 589                 inc_node_page_state(page, NR_ISOLATED_ANON + 0);
 590                 VM_BUG_ON_PAGE(!PageLocked(page), page);
 591                 VM_BUG_ON_PAGE(PageLRU(page), page);
 592
 593                 /* There should be enough young pte to collapse the page */
 594                 if (pte_young(pteval) ||
 595                     page_is_young(page) || PageReferenced(page) ||
 596                     mmu_notifier_test_young(vma->vm_mm, address))
 597                         referenced++;
 598         }
 599
 600         if (unlikely(!writable)) {
 601                 result = SCAN_PAGE_RO;
 602         } else if (unlikely(!referenced)) {
 603                 result = SCAN_LACK_REFERENCED_PAGE;
 604         } else {
 605                 result = SCAN_SUCCEED;
 606                 trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 607                                                     referenced, writable, result);
 608                 return 1;
 609         }
 610 out:
 611         release_pte_pages(pte, _pte);
 612         trace_mm_collapse_huge_page_isolate(page, none_or_zero,
 613                                             referenced, writable, result);
 614         return 0;
 615 }
 616
 617 static void __collapse_huge_page_copy(pte_t *pte, struct page *page,
 618                                       struct vm_area_struct *vma,
 619                                       unsigned long address,
 620                                       spinlock_t *ptl)
 621 {
 622         pte_t *_pte;
 623         for (_pte = pte; _pte < pte+HPAGE_PMD_NR; _pte++) {
 624                 pte_t pteval = *_pte;
 625                 struct page *src_page;
 626
 627                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
 628                         clear_user_highpage(page, address);
 629                         add_mm_counter(vma->vm_mm, MM_ANONPAGES, 1);
 630                         if (is_zero_pfn(pte_pfn(pteval))) {
 631                                 /*
 632                                  * ptl mostly unnecessary.
 633                                  */
 634                                 spin_lock(ptl);
 635                                 /*
 636                                  * paravirt calls inside pte_clear here are
 637                                  * superfluous.
 638                                  */
 639                                 pte_clear(vma->vm_mm, address, _pte);
 640                                 spin_unlock(ptl);
 641                         }
 642                 } else {
 643                         src_page = pte_page(pteval);
 644                         copy_user_highpage(page, src_page, address, vma);
 645                         VM_BUG_ON_PAGE(page_mapcount(src_page) != 1, src_page);
 646                         release_pte_page(src_page);
 647                         /*
 648                          * ptl mostly unnecessary, but preempt has to
 649                          * be disabled to update the per-cpu stats
 650                          * inside page_remove_rmap().
 651                          */
 652                         spin_lock(ptl);
 653                         /*
 654                          * paravirt calls inside pte_clear here are
 655                          * superfluous.
 656                          */
 657                         pte_clear(vma->vm_mm, address, _pte);
 658                         page_remove_rmap(src_page, false);
 659                         spin_unlock(ptl);
 660                         free_page_and_swap_cache(src_page);
 661                 }
 662
 663                 address += PAGE_SIZE;
 664                 page++;
 665         }
 666 }
 667
 668 static void khugepaged_alloc_sleep(void)
 669 {
 670         DEFINE_WAIT(wait);
 671
 672         add_wait_queue(&khugepaged_wait, &wait);
 673         freezable_schedule_timeout_interruptible(
 674                 msecs_to_jiffies(khugepaged_alloc_sleep_millisecs));
 675         remove_wait_queue(&khugepaged_wait, &wait);
 676 }
 677
 678 static int khugepaged_node_load[MAX_NUMNODES];
 679
 680 static bool khugepaged_scan_abort(int nid)
 681 {
 682         int i;
 683
 684         /*
 685          * If node_reclaim_mode is disabled, then no extra effort is made to
 686          * allocate memory locally.
 687          */
 688         if (!node_reclaim_mode)
 689                 return false;
 690
 691         /* If there is a count for this node already, it must be acceptable */
 692         if (khugepaged_node_load[nid])
 693                 return false;
 694
 695         for (i = 0; i < MAX_NUMNODES; i++) {
 696                 if (!khugepaged_node_load[i])
 697                         continue;
 698                 if (node_distance(nid, i) > RECLAIM_DISTANCE)
 699                         return true;
 700         }
 701         return false;
 702 }
 703
 704 /* Defrag for khugepaged will enter direct reclaim/compaction if necessary */
 705 static inline gfp_t alloc_hugepage_khugepaged_gfpmask(void)
 706 {
 707         return khugepaged_defrag() ? GFP_TRANSHUGE : GFP_TRANSHUGE_LIGHT;
 708 }
 709
 710 #ifdef CONFIG_NUMA
 711 static int khugepaged_find_target_node(void)
 712 {
 713         static int last_khugepaged_target_node = NUMA_NO_NODE;
 714         int nid, target_node = 0, max_value = 0;
 715
 716         /* find first node with max normal pages hit */
 717         for (nid = 0; nid < MAX_NUMNODES; nid++)
 718                 if (khugepaged_node_load[nid] > max_value) {
 719                         max_value = khugepaged_node_load[nid];
 720                         target_node = nid;
 721                 }
 722
 723         /* do some balance if several nodes have the same hit record */
 724         if (target_node <= last_khugepaged_target_node)
 725                 for (nid = last_khugepaged_target_node + 1; nid < MAX_NUMNODES;
 726                                 nid++)
 727                         if (max_value == khugepaged_node_load[nid]) {
 728                                 target_node = nid;
 729                                 break;
 730                         }
 731
 732         last_khugepaged_target_node = target_node;
 733         return target_node;
 734 }
 735
 736 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 737 {
 738         if (IS_ERR(*hpage)) {
 739                 if (!*wait)
 740                         return false;
 741
 742                 *wait = false;
 743                 *hpage = NULL;
 744                 khugepaged_alloc_sleep();
 745         } else if (*hpage) {
 746                 put_page(*hpage);
 747                 *hpage = NULL;
 748         }
 749
 750         return true;
 751 }
 752
 753 static struct page *
 754 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 755 {
 756         VM_BUG_ON_PAGE(*hpage, *hpage);
 757
 758         *hpage = __alloc_pages_node(node, gfp, HPAGE_PMD_ORDER);
 759         if (unlikely(!*hpage)) {
 760                 count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 761                 *hpage = ERR_PTR(-ENOMEM);
 762                 return NULL;
 763         }
 764
 765         prep_transhuge_page(*hpage);
 766         count_vm_event(THP_COLLAPSE_ALLOC);
 767         return *hpage;
 768 }
 769 #else
 770 static int khugepaged_find_target_node(void)
 771 {
 772         return 0;
 773 }
 774
 775 static inline struct page *alloc_khugepaged_hugepage(void)
 776 {
 777         struct page *page;
 778
 779         page = alloc_pages(alloc_hugepage_khugepaged_gfpmask(),
 780                            HPAGE_PMD_ORDER);
 781         if (page)
 782                 prep_transhuge_page(page);
 783         return page;
 784 }
 785
 786 static struct page *khugepaged_alloc_hugepage(bool *wait)
 787 {
 788         struct page *hpage;
 789
 790         do {
 791                 hpage = alloc_khugepaged_hugepage();
 792                 if (!hpage) {
 793                         count_vm_event(THP_COLLAPSE_ALLOC_FAILED);
 794                         if (!*wait)
 795                                 return NULL;
 796
 797                         *wait = false;
 798                         khugepaged_alloc_sleep();
 799                 } else
 800                         count_vm_event(THP_COLLAPSE_ALLOC);
 801         } while (unlikely(!hpage) && likely(khugepaged_enabled()));
 802
 803         return hpage;
 804 }
 805
 806 static bool khugepaged_prealloc_page(struct page **hpage, bool *wait)
 807 {
 808         /*
 809          * If the hpage allocated earlier was briefly exposed in page cache
 810          * before collapse_file() failed, it is possible that racing lookups
 811          * have not yet completed, and would then be unpleasantly surprised by
 812          * finding the hpage reused for the same mapping at a different offset.
 813          * Just release the previous allocation if there is any danger of that.
 814          */
 815         if (*hpage && page_count(*hpage) > 1) {
 816                 put_page(*hpage);
 817                 *hpage = NULL;
 818         }
 819
 820         if (!*hpage)
 821                 *hpage = khugepaged_alloc_hugepage(wait);
 822
 823         if (unlikely(!*hpage))
 824                 return false;
 825
 826         return true;
 827 }
 828
 829 static struct page *
 830 khugepaged_alloc_page(struct page **hpage, gfp_t gfp, int node)
 831 {
 832         VM_BUG_ON(!*hpage);
 833
 834         return  *hpage;
 835 }
 836 #endif
 837
 838 static bool hugepage_vma_check(struct vm_area_struct *vma)
 839 {
 840         if ((!(vma->vm_flags & VM_HUGEPAGE) && !khugepaged_always()) ||
 841             (vma->vm_flags & VM_NOHUGEPAGE))
 842                 return false;
 843         if (shmem_file(vma->vm_file)) {
 844                 if (!IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE))
 845                         return false;
 846                 return IS_ALIGNED((vma->vm_start >> PAGE_SHIFT) - vma->vm_pgoff,
 847                                 HPAGE_PMD_NR);
 848         }
 849         if (!vma->anon_vma || vma->vm_ops)
 850                 return false;
 851         if (is_vma_temporary_stack(vma))
 852                 return false;
 853         return !(vma->vm_flags & VM_NO_KHUGEPAGED);
 854 }
 855
 856 /*
 857  * If mmap_sem temporarily dropped, revalidate vma
 858  * before taking mmap_sem.
 859  * Return 0 if succeeds, otherwise return none-zero
 860  * value (scan code).
 861  */
 862
 863 static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address,
 864                 struct vm_area_struct **vmap)
 865 {
 866         struct vm_area_struct *vma;
 867         unsigned long hstart, hend;
 868
 869         if (unlikely(khugepaged_test_exit(mm)))
 870                 return SCAN_ANY_PROCESS;
 871
 872         *vmap = vma = find_vma(mm, address);
 873         if (!vma)
 874                 return SCAN_VMA_NULL;
 875
 876         hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
 877         hend = vma->vm_end & HPAGE_PMD_MASK;
 878         if (address < hstart || address + HPAGE_PMD_SIZE > hend)
 879                 return SCAN_ADDRESS_RANGE;
 880         if (!hugepage_vma_check(vma))
 881                 return SCAN_VMA_CHECK;
 882         return 0;
 883 }
 884
 885 /*
 886  * Bring missing pages in from swap, to complete THP collapse.
 887  * Only done if khugepaged_scan_pmd believes it is worthwhile.
 888  *
 889  * Called and returns without pte mapped or spinlocks held,
 890  * but with mmap_sem held to protect against vma changes.
 891  */
 892
 893 static bool __collapse_huge_page_swapin(struct mm_struct *mm,
 894                                         struct vm_area_struct *vma,
 895                                         unsigned long address, pmd_t *pmd,
 896                                         int referenced)
 897 {
 898         pte_t pteval;
 899         int swapped_in = 0, ret = 0;
 900         struct fault_env fe = {
 901                 .vma = vma,
 902                 .address = address,
 903                 .flags = FAULT_FLAG_ALLOW_RETRY,
 904                 .pmd = pmd,
 905         };
 906
 907         /* we only decide to swapin, if there is enough young ptes */
 908         if (referenced < HPAGE_PMD_NR/2) {
 909                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 910                 return false;
 911         }
 912         fe.pte = pte_offset_map(pmd, address);
 913         for (; fe.address < address + HPAGE_PMD_NR*PAGE_SIZE;
 914                         fe.pte++, fe.address += PAGE_SIZE) {
 915                 pteval = *fe.pte;
 916                 if (!is_swap_pte(pteval))
 917                         continue;
 918                 swapped_in++;
 919                 ret = do_swap_page(&fe, pteval);
 920
 921                 /* do_swap_page returns VM_FAULT_RETRY with released mmap_sem */
 922                 if (ret & VM_FAULT_RETRY) {
 923                         down_read(&mm->mmap_sem);
 924                         if (hugepage_vma_revalidate(mm, address, &fe.vma)) {
 925                                 /* vma is no longer available, don't continue to swapin */
 926                                 trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 927                                 return false;
 928                         }
 929                         /* check if the pmd is still valid */
 930                         if (mm_find_pmd(mm, address) != pmd)
 931                                 return false;
 932                 }
 933                 if (ret & VM_FAULT_ERROR) {
 934                         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 0);
 935                         return false;
 936                 }
 937                 /* pte is unmapped now, we need to map it */
 938                 fe.pte = pte_offset_map(pmd, fe.address);
 939         }
 940         fe.pte--;
 941         pte_unmap(fe.pte);
 942         trace_mm_collapse_huge_page_swapin(mm, swapped_in, referenced, 1);
 943         return true;
 944 }
 945
 946 static void collapse_huge_page(struct mm_struct *mm,
 947                                    unsigned long address,
 948                                    struct page **hpage,
 949                                    int node, int referenced)
 950 {
 951         pmd_t *pmd, _pmd;
 952         pte_t *pte;
 953         pgtable_t pgtable;
 954         struct page *new_page;
 955         spinlock_t *pmd_ptl, *pte_ptl;
 956         int isolated = 0, result = 0;
 957         struct mem_cgroup *memcg;
 958         struct vm_area_struct *vma;
 959         unsigned long mmun_start;       /* For mmu_notifiers */
 960         unsigned long mmun_end;         /* For mmu_notifiers */
 961         gfp_t gfp;
 962
 963         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
 964
 965         /* Only allocate from the target node */
 966         gfp = alloc_hugepage_khugepaged_gfpmask() | __GFP_OTHER_NODE | __GFP_THISNODE;
 967
 968         /*
 969          * Before allocating the hugepage, release the mmap_sem read lock.
 970          * The allocation can take potentially a long time if it involves
 971          * sync compaction, and we do not need to hold the mmap_sem during
 972          * that. We will recheck the vma after taking it again in write mode.
 973          */
 974         up_read(&mm->mmap_sem);
 975         new_page = khugepaged_alloc_page(hpage, gfp, node);
 976         if (!new_page) {
 977                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
 978                 goto out_nolock;
 979         }
 980
 981         /* Do not oom kill for khugepaged charges */
 982         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp | __GFP_NORETRY,
 983                                            &memcg, true))) {
 984                 result = SCAN_CGROUP_CHARGE_FAIL;
 985                 goto out_nolock;
 986         }
 987
 988         down_read(&mm->mmap_sem);
 989         result = hugepage_vma_revalidate(mm, address, &vma);
 990         if (result) {
 991                 mem_cgroup_cancel_charge(new_page, memcg, true);
 992                 up_read(&mm->mmap_sem);
 993                 goto out_nolock;
 994         }
 995
 996         pmd = mm_find_pmd(mm, address);
 997         if (!pmd) {
 998                 result = SCAN_PMD_NULL;
 999                 mem_cgroup_cancel_charge(new_page, memcg, true);
1000                 up_read(&mm->mmap_sem);
1001                 goto out_nolock;
1002         }
1003
1004         /*
1005          * __collapse_huge_page_swapin always returns with mmap_sem locked.
1006          * If it fails, we release mmap_sem and jump out_nolock.
1007          * Continuing to collapse causes inconsistency.
1008          */
1009         if (!__collapse_huge_page_swapin(mm, vma, address, pmd, referenced)) {
1010                 mem_cgroup_cancel_charge(new_page, memcg, true);
1011                 up_read(&mm->mmap_sem);
1012                 goto out_nolock;
1013         }
1014
1015         up_read(&mm->mmap_sem);
1016         /*
1017          * Prevent all access to pagetables with the exception of
1018          * gup_fast later handled by the ptep_clear_flush and the VM
1019          * handled by the anon_vma lock + PG_lock.
1020          */
1021         down_write(&mm->mmap_sem);
1022         result = hugepage_vma_revalidate(mm, address, &vma);
1023         if (result)
1024                 goto out;
1025         /* check if the pmd is still valid */
1026         if (mm_find_pmd(mm, address) != pmd)
1027                 goto out;
1028
1029         anon_vma_lock_write(vma->anon_vma);
1030
1031         pte = pte_offset_map(pmd, address);
1032         pte_ptl = pte_lockptr(mm, pmd);
1033
1034         mmun_start = address;
1035         mmun_end   = address + HPAGE_PMD_SIZE;
1036         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
1037         pmd_ptl = pmd_lock(mm, pmd); /* probably unnecessary */
1038         /*
1039          * After this gup_fast can't run anymore. This also removes
1040          * any huge TLB entry from the CPU so we won't allow
1041          * huge and small TLB entries for the same virtual address
1042          * to avoid the risk of CPU bugs in that area.
1043          */
1044         _pmd = pmdp_collapse_flush(vma, address, pmd);
1045         spin_unlock(pmd_ptl);
1046         mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
1047
1048         spin_lock(pte_ptl);
1049         isolated = __collapse_huge_page_isolate(vma, address, pte);
1050         spin_unlock(pte_ptl);
1051
1052         if (unlikely(!isolated)) {
1053                 pte_unmap(pte);
1054                 spin_lock(pmd_ptl);
1055                 BUG_ON(!pmd_none(*pmd));
1056                 /*
1057                  * We can only use set_pmd_at when establishing
1058                  * hugepmds and never for establishing regular pmds that
1059                  * points to regular pagetables. Use pmd_populate for that
1060                  */
1061                 pmd_populate(mm, pmd, pmd_pgtable(_pmd));
1062                 spin_unlock(pmd_ptl);
1063                 anon_vma_unlock_write(vma->anon_vma);
1064                 result = SCAN_FAIL;
1065                 goto out;
1066         }
1067
1068         /*
1069          * All pages are isolated and locked so anon_vma rmap
1070          * can't run anymore.
1071          */
1072         anon_vma_unlock_write(vma->anon_vma);
1073
1074         __collapse_huge_page_copy(pte, new_page, vma, address, pte_ptl);
1075         pte_unmap(pte);
1076         __SetPageUptodate(new_page);
1077         pgtable = pmd_pgtable(_pmd);
1078
1079         _pmd = mk_huge_pmd(new_page, vma->vm_page_prot);
1080         _pmd = maybe_pmd_mkwrite(pmd_mkdirty(_pmd), vma);
1081
1082         /*
1083          * spin_lock() below is not the equivalent of smp_wmb(), so
1084          * this is needed to avoid the copy_huge_page writes to become
1085          * visible after the set_pmd_at() write.
1086          */
1087         smp_wmb();
1088
1089         spin_lock(pmd_ptl);
1090         BUG_ON(!pmd_none(*pmd));
1091         page_add_new_anon_rmap(new_page, vma, address, true);
1092         mem_cgroup_commit_charge(new_page, memcg, false, true);
1093         lru_cache_add_active_or_unevictable(new_page, vma);
1094         pgtable_trans_huge_deposit(mm, pmd, pgtable);
1095         set_pmd_at(mm, address, pmd, _pmd);
1096         update_mmu_cache_pmd(vma, address, pmd);
1097         spin_unlock(pmd_ptl);
1098
1099         *hpage = NULL;
1100
1101         khugepaged_pages_collapsed++;
1102         result = SCAN_SUCCEED;
1103 out_up_write:
1104         up_write(&mm->mmap_sem);
1105 out_nolock:
1106         trace_mm_collapse_huge_page(mm, isolated, result);
1107         return;
1108 out:
1109         mem_cgroup_cancel_charge(new_page, memcg, true);
1110         goto out_up_write;
1111 }
1112
1113 static int khugepaged_scan_pmd(struct mm_struct *mm,
1114                                struct vm_area_struct *vma,
1115                                unsigned long address,
1116                                struct page **hpage)
1117 {
1118         pmd_t *pmd;
1119         pte_t *pte, *_pte;
1120         int ret = 0, none_or_zero = 0, result = 0, referenced = 0;
1121         struct page *page = NULL;
1122         unsigned long _address;
1123         spinlock_t *ptl;
1124         int node = NUMA_NO_NODE, unmapped = 0;
1125         bool writable = false;
1126
1127         VM_BUG_ON(address & ~HPAGE_PMD_MASK);
1128
1129         pmd = mm_find_pmd(mm, address);
1130         if (!pmd) {
1131                 result = SCAN_PMD_NULL;
1132                 goto out;
1133         }
1134
1135         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1136         pte = pte_offset_map_lock(mm, pmd, address, &ptl);
1137         for (_address = address, _pte = pte; _pte < pte+HPAGE_PMD_NR;
1138              _pte++, _address += PAGE_SIZE) {
1139                 pte_t pteval = *_pte;
1140                 if (is_swap_pte(pteval)) {
1141                         if (++unmapped <= khugepaged_max_ptes_swap) {
1142                                 continue;
1143                         } else {
1144                                 result = SCAN_EXCEED_SWAP_PTE;
1145                                 goto out_unmap;
1146                         }
1147                 }
1148                 if (pte_none(pteval) || is_zero_pfn(pte_pfn(pteval))) {
1149                         if (!userfaultfd_armed(vma) &&
1150                             ++none_or_zero <= khugepaged_max_ptes_none) {
1151                                 continue;
1152                         } else {
1153                                 result = SCAN_EXCEED_NONE_PTE;
1154                                 goto out_unmap;
1155                         }
1156                 }
1157                 if (!pte_present(pteval)) {
1158                         result = SCAN_PTE_NON_PRESENT;
1159                         goto out_unmap;
1160                 }
1161                 if (pte_write(pteval))
1162                         writable = true;
1163
1164                 page = vm_normal_page(vma, _address, pteval);
1165                 if (unlikely(!page)) {
1166                         result = SCAN_PAGE_NULL;
1167                         goto out_unmap;
1168                 }
1169
1170                 /* TODO: teach khugepaged to collapse THP mapped with pte */
1171                 if (PageCompound(page)) {
1172                         result = SCAN_PAGE_COMPOUND;
1173                         goto out_unmap;
1174                 }
1175
1176                 /*
1177                  * Record which node the original page is from and save this
1178                  * information to khugepaged_node_load[].
1179                  * Khupaged will allocate hugepage from the node has the max
1180                  * hit record.
1181                  */
1182                 node = page_to_nid(page);
1183                 if (khugepaged_scan_abort(node)) {
1184                         result = SCAN_SCAN_ABORT;
1185                         goto out_unmap;
1186                 }
1187                 khugepaged_node_load[node]++;
1188                 if (!PageLRU(page)) {
1189                         result = SCAN_PAGE_LRU;
1190                         goto out_unmap;
1191                 }
1192                 if (PageLocked(page)) {
1193                         result = SCAN_PAGE_LOCK;
1194                         goto out_unmap;
1195                 }
1196                 if (!PageAnon(page)) {
1197                         result = SCAN_PAGE_ANON;
1198                         goto out_unmap;
1199                 }
1200
1201                 /*
1202                  * cannot use mapcount: can't collapse if there's a gup pin.
1203                  * The page must only be referenced by the scanned process
1204                  * and page swap cache.
1205                  */
1206                 if (page_count(page) != 1 + !!PageSwapCache(page)) {
1207                         result = SCAN_PAGE_COUNT;
1208                         goto out_unmap;
1209                 }
1210                 if (pte_young(pteval) ||
1211                     page_is_young(page) || PageReferenced(page) ||
1212                     mmu_notifier_test_young(vma->vm_mm, address))
1213                         referenced++;
1214         }
1215         if (writable) {
1216                 if (referenced) {
1217                         result = SCAN_SUCCEED;
1218                         ret = 1;
1219                 } else {
1220                         result = SCAN_LACK_REFERENCED_PAGE;
1221                 }
1222         } else {
1223                 result = SCAN_PAGE_RO;
1224         }
1225 out_unmap:
1226         pte_unmap_unlock(pte, ptl);
1227         if (ret) {
1228                 node = khugepaged_find_target_node();
1229                 /* collapse_huge_page will return with the mmap_sem released */
1230                 collapse_huge_page(mm, address, hpage, node, referenced);
1231         }
1232 out:
1233         trace_mm_khugepaged_scan_pmd(mm, page, writable, referenced,
1234                                      none_or_zero, result, unmapped);
1235         return ret;
1236 }
1237
1238 static void collect_mm_slot(struct mm_slot *mm_slot)
1239 {
1240         struct mm_struct *mm = mm_slot->mm;
1241
1242         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1243
1244         if (khugepaged_test_exit(mm)) {
1245                 /* free mm_slot */
1246                 hash_del(&mm_slot->hash);
1247                 list_del(&mm_slot->mm_node);
1248
1249                 /*
1250                  * Not strictly needed because the mm exited already.
1251                  *
1252                  * clear_bit(MMF_VM_HUGEPAGE, &mm->flags);
1253                  */
1254
1255                 /* khugepaged_mm_lock actually not necessary for the below */
1256                 free_mm_slot(mm_slot);
1257                 mmdrop(mm);
1258         }
1259 }
1260
1261 #if defined(CONFIG_SHMEM) && defined(CONFIG_TRANSPARENT_HUGE_PAGECACHE)
1262 static void retract_page_tables(struct address_space *mapping, pgoff_t pgoff)
1263 {
1264         struct vm_area_struct *vma;
1265         struct mm_struct *mm;
1266         unsigned long addr;
1267         pmd_t *pmd, _pmd;
1268
1269         i_mmap_lock_write(mapping);
1270         vma_interval_tree_foreach(vma, &mapping->i_mmap, pgoff, pgoff) {
1271                 /* probably overkill */
1272                 if (vma->anon_vma)
1273                         continue;
1274                 addr = vma->vm_start + ((pgoff - vma->vm_pgoff) << PAGE_SHIFT);
1275                 if (addr & ~HPAGE_PMD_MASK)
1276                         continue;
1277                 if (vma->vm_end < addr + HPAGE_PMD_SIZE)
1278                         continue;
1279                 mm = vma->vm_mm;
1280                 pmd = mm_find_pmd(mm, addr);
1281                 if (!pmd)
1282                         continue;
1283                 /*
1284                  * We need exclusive mmap_sem to retract page table.
1285                  * If trylock fails we would end up with pte-mapped THP after
1286                  * re-fault. Not ideal, but it's more important to not disturb
1287                  * the system too much.
1288                  */
1289                 if (down_write_trylock(&mm->mmap_sem)) {
1290                         if (!khugepaged_test_exit(mm)) {
1291                                 spinlock_t *ptl = pmd_lock(mm, pmd);
1292                                 /* assume page table is clear */
1293                                 _pmd = pmdp_collapse_flush(vma, addr, pmd);
1294                                 spin_unlock(ptl);
1295                                 atomic_long_dec(&mm->nr_ptes);
1296                                 pte_free(mm, pmd_pgtable(_pmd));
1297                         }
1298                         up_write(&mm->mmap_sem);
1299                 }
1300         }
1301         i_mmap_unlock_write(mapping);
1302 }
1303
1304 /**
1305  * collapse_shmem - collapse small tmpfs/shmem pages into huge one.
1306  *
1307  * Basic scheme is simple, details are more complex:
1308  *  - allocate and lock a new huge page;
1309  *  - scan over radix tree replacing old pages the new one
1310  *    + swap in pages if necessary;
1311  *    + fill in gaps;
1312  *    + keep old pages around in case if rollback is required;
1313  *  - if replacing succeed:
1314  *    + copy data over;
1315  *    + free old pages;
1316  *    + unlock huge page;
1317  *  - if replacing failed;
1318  *    + put all pages back and unfreeze them;
1319  *    + restore gaps in the radix-tree;
1320  *    + unlock and free huge page;
1321  */
1322 static void collapse_shmem(struct mm_struct *mm,
1323                 struct address_space *mapping, pgoff_t start,
1324                 struct page **hpage, int node)
1325 {
1326         gfp_t gfp;
1327         struct page *page, *new_page, *tmp;
1328         struct mem_cgroup *memcg;
1329         pgoff_t index, end = start + HPAGE_PMD_NR;
1330         LIST_HEAD(pagelist);
1331         struct radix_tree_iter iter;
1332         void **slot;
1333         int nr_none = 0, result = SCAN_SUCCEED;
1334
1335         VM_BUG_ON(start & (HPAGE_PMD_NR - 1));
1336
1337         /* Only allocate from the target node */
1338         gfp = alloc_hugepage_khugepaged_gfpmask() |
1339                 __GFP_OTHER_NODE | __GFP_THISNODE;
1340
1341         new_page = khugepaged_alloc_page(hpage, gfp, node);
1342         if (!new_page) {
1343                 result = SCAN_ALLOC_HUGE_PAGE_FAIL;
1344                 goto out;
1345         }
1346
1347         /* Do not oom kill for khugepaged charges */
1348         if (unlikely(mem_cgroup_try_charge(new_page, mm, gfp | __GFP_NORETRY,
1349                                            &memcg, true))) {
1350                 result = SCAN_CGROUP_CHARGE_FAIL;
1351                 goto out;
1352         }
1353
1354         __SetPageLocked(new_page);
1355         __SetPageSwapBacked(new_page);
1356         new_page->index = start;
1357         new_page->mapping = mapping;
1358
1359         /*
1360          * At this point the new_page is locked and not up-to-date.
1361          * It's safe to insert it into the page cache, because nobody would
1362          * be able to map it or use it in another way until we unlock it.
1363          */
1364
1365         index = start;
1366         spin_lock_irq(&mapping->tree_lock);
1367         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1368                 int n = min(iter.index, end) - index;
1369
1370                 /*
1371                  * Stop if extent has been hole-punched, and is now completely
1372                  * empty (the more obvious i_size_read() check would take an
1373                  * irq-unsafe seqlock on 32-bit).
1374                  */
1375                 if (n >= HPAGE_PMD_NR) {
1376                         result = SCAN_TRUNCATED;
1377                         goto tree_locked;
1378                 }
1379
1380                 /*
1381                  * Handle holes in the radix tree: charge it from shmem and
1382                  * insert relevant subpage of new_page into the radix-tree.
1383                  */
1384                 if (n && !shmem_charge(mapping->host, n)) {
1385                         result = SCAN_FAIL;
1386                         goto tree_locked;
1387                 }
1388                 for (; index < min(iter.index, end); index++) {
1389                         radix_tree_insert(&mapping->page_tree, index,
1390                                         new_page + (index % HPAGE_PMD_NR));
1391                 }
1392                 nr_none += n;
1393
1394                 /* We are done. */
1395                 if (index >= end)
1396                         break;
1397
1398                 page = radix_tree_deref_slot_protected(slot,
1399                                 &mapping->tree_lock);
1400                 if (radix_tree_exceptional_entry(page) || !PageUptodate(page)) {
1401                         spin_unlock_irq(&mapping->tree_lock);
1402                         /* swap in or instantiate fallocated page */
1403                         if (shmem_getpage(mapping->host, index, &page,
1404                                                 SGP_NOHUGE)) {
1405                                 result = SCAN_FAIL;
1406                                 goto tree_unlocked;
1407                         }
1408                 } else if (trylock_page(page)) {
1409                         get_page(page);
1410                         spin_unlock_irq(&mapping->tree_lock);
1411                 } else {
1412                         result = SCAN_PAGE_LOCK;
1413                         goto tree_locked;
1414                 }
1415
1416                 /*
1417                  * The page must be locked, so we can drop the tree_lock
1418                  * without racing with truncate.
1419                  */
1420                 VM_BUG_ON_PAGE(!PageLocked(page), page);
1421                 VM_BUG_ON_PAGE(!PageUptodate(page), page);
1422
1423                 /*
1424                  * If file was truncated then extended, or hole-punched, before
1425                  * we locked the first page, then a THP might be there already.
1426                  */
1427                 if (PageTransCompound(page)) {
1428                         result = SCAN_PAGE_COMPOUND;
1429                         goto out_unlock;
1430                 }
1431
1432                 if (page_mapping(page) != mapping) {
1433                         result = SCAN_TRUNCATED;
1434                         goto out_unlock;
1435                 }
1436
1437                 if (isolate_lru_page(page)) {
1438                         result = SCAN_DEL_PAGE_LRU;
1439                         goto out_unlock;
1440                 }
1441
1442                 if (page_mapped(page))
1443                         unmap_mapping_range(mapping, index << PAGE_SHIFT,
1444                                         PAGE_SIZE, 0);
1445
1446                 spin_lock_irq(&mapping->tree_lock);
1447
1448                 slot = radix_tree_lookup_slot(&mapping->page_tree, index);
1449                 VM_BUG_ON_PAGE(page != radix_tree_deref_slot_protected(slot,
1450                                         &mapping->tree_lock), page);
1451                 VM_BUG_ON_PAGE(page_mapped(page), page);
1452
1453                 /*
1454                  * The page is expected to have page_count() == 3:
1455                  *  - we hold a pin on it;
1456                  *  - one reference from radix tree;
1457                  *  - one from isolate_lru_page;
1458                  */
1459                 if (!page_ref_freeze(page, 3)) {
1460                         result = SCAN_PAGE_COUNT;
1461                         spin_unlock_irq(&mapping->tree_lock);
1462                         putback_lru_page(page);
1463                         goto out_unlock;
1464                 }
1465
1466                 /*
1467                  * Add the page to the list to be able to undo the collapse if
1468                  * something go wrong.
1469                  */
1470                 list_add_tail(&page->lru, &pagelist);
1471
1472                 /* Finally, replace with the new page. */
1473                 radix_tree_replace_slot(slot,
1474                                 new_page + (index % HPAGE_PMD_NR));
1475
1476                 slot = radix_tree_iter_next(&iter);
1477                 index++;
1478                 continue;
1479 out_unlock:
1480                 unlock_page(page);
1481                 put_page(page);
1482                 goto tree_unlocked;
1483         }
1484
1485         /*
1486          * Handle hole in radix tree at the end of the range.
1487          * This code only triggers if there's nothing in radix tree
1488          * beyond 'end'.
1489          */
1490         if (index < end) {
1491                 int n = end - index;
1492
1493                 /* Stop if extent has been truncated, and is now empty */
1494                 if (n >= HPAGE_PMD_NR) {
1495                         result = SCAN_TRUNCATED;
1496                         goto tree_locked;
1497                 }
1498                 if (!shmem_charge(mapping->host, n)) {
1499                         result = SCAN_FAIL;
1500                         goto tree_locked;
1501                 }
1502                 for (; index < end; index++) {
1503                         radix_tree_insert(&mapping->page_tree, index,
1504                                         new_page + (index % HPAGE_PMD_NR));
1505                 }
1506                 nr_none += n;
1507         }
1508
1509         __inc_node_page_state(new_page, NR_SHMEM_THPS);
1510         if (nr_none) {
1511                 struct zone *zone = page_zone(new_page);
1512
1513                 __mod_node_page_state(zone->zone_pgdat, NR_FILE_PAGES, nr_none);
1514                 __mod_node_page_state(zone->zone_pgdat, NR_SHMEM, nr_none);
1515         }
1516
1517 tree_locked:
1518         spin_unlock_irq(&mapping->tree_lock);
1519 tree_unlocked:
1520
1521         if (result == SCAN_SUCCEED) {
1522                 /*
1523                  * Replacing old pages with new one has succeed, now we need to
1524                  * copy the content and free old pages.
1525                  */
1526                 index = start;
1527                 list_for_each_entry_safe(page, tmp, &pagelist, lru) {
1528                         while (index < page->index) {
1529                                 clear_highpage(new_page + (index % HPAGE_PMD_NR));
1530                                 index++;
1531                         }
1532                         copy_highpage(new_page + (page->index % HPAGE_PMD_NR),
1533                                         page);
1534                         list_del(&page->lru);
1535                         page->mapping = NULL;
1536                         page_ref_unfreeze(page, 1);
1537                         ClearPageActive(page);
1538                         ClearPageUnevictable(page);
1539                         unlock_page(page);
1540                         put_page(page);
1541                         index++;
1542                 }
1543                 while (index < end) {
1544                         clear_highpage(new_page + (index % HPAGE_PMD_NR));
1545                         index++;
1546                 }
1547
1548                 SetPageUptodate(new_page);
1549                 page_ref_add(new_page, HPAGE_PMD_NR - 1);
1550                 set_page_dirty(new_page);
1551                 mem_cgroup_commit_charge(new_page, memcg, false, true);
1552                 lru_cache_add_anon(new_page);
1553
1554                 /*
1555                  * Remove pte page tables, so we can re-fault the page as huge.
1556                  */
1557                 retract_page_tables(mapping, start);
1558                 *hpage = NULL;
1559         } else {
1560                 /* Something went wrong: rollback changes to the radix-tree */
1561                 spin_lock_irq(&mapping->tree_lock);
1562                 mapping->nrpages -= nr_none;
1563                 shmem_uncharge(mapping->host, nr_none);
1564
1565                 radix_tree_for_each_slot(slot, &mapping->page_tree, &iter,
1566                                 start) {
1567                         if (iter.index >= end)
1568                                 break;
1569                         page = list_first_entry_or_null(&pagelist,
1570                                         struct page, lru);
1571                         if (!page || iter.index < page->index) {
1572                                 if (!nr_none)
1573                                         break;
1574                                 nr_none--;
1575                                 /* Put holes back where they were */
1576                                 radix_tree_delete(&mapping->page_tree,
1577                                                   iter.index);
1578                                 slot = radix_tree_iter_next(&iter);
1579                                 continue;
1580                         }
1581
1582                         VM_BUG_ON_PAGE(page->index != iter.index, page);
1583
1584                         /* Unfreeze the page. */
1585                         list_del(&page->lru);
1586                         page_ref_unfreeze(page, 2);
1587                         radix_tree_replace_slot(slot, page);
1588                         spin_unlock_irq(&mapping->tree_lock);
1589                         unlock_page(page);
1590                         putback_lru_page(page);
1591                         spin_lock_irq(&mapping->tree_lock);
1592                         slot = radix_tree_iter_next(&iter);
1593                 }
1594                 VM_BUG_ON(nr_none);
1595                 spin_unlock_irq(&mapping->tree_lock);
1596
1597                 mem_cgroup_cancel_charge(new_page, memcg, true);
1598                 new_page->mapping = NULL;
1599         }
1600
1601         unlock_page(new_page);
1602 out:
1603         VM_BUG_ON(!list_empty(&pagelist));
1604         /* TODO: tracepoints */
1605 }
1606
1607 static void khugepaged_scan_shmem(struct mm_struct *mm,
1608                 struct address_space *mapping,
1609                 pgoff_t start, struct page **hpage)
1610 {
1611         struct page *page = NULL;
1612         struct radix_tree_iter iter;
1613         void **slot;
1614         int present, swap;
1615         int node = NUMA_NO_NODE;
1616         int result = SCAN_SUCCEED;
1617
1618         present = 0;
1619         swap = 0;
1620         memset(khugepaged_node_load, 0, sizeof(khugepaged_node_load));
1621         rcu_read_lock();
1622         radix_tree_for_each_slot(slot, &mapping->page_tree, &iter, start) {
1623                 if (iter.index >= start + HPAGE_PMD_NR)
1624                         break;
1625
1626                 page = radix_tree_deref_slot(slot);
1627                 if (radix_tree_deref_retry(page)) {
1628                         slot = radix_tree_iter_retry(&iter);
1629                         continue;
1630                 }
1631
1632                 if (radix_tree_exception(page)) {
1633                         if (++swap > khugepaged_max_ptes_swap) {
1634                                 result = SCAN_EXCEED_SWAP_PTE;
1635                                 break;
1636                         }
1637                         continue;
1638                 }
1639
1640                 if (PageTransCompound(page)) {
1641                         result = SCAN_PAGE_COMPOUND;
1642                         break;
1643                 }
1644
1645                 node = page_to_nid(page);
1646                 if (khugepaged_scan_abort(node)) {
1647                         result = SCAN_SCAN_ABORT;
1648                         break;
1649                 }
1650                 khugepaged_node_load[node]++;
1651
1652                 if (!PageLRU(page)) {
1653                         result = SCAN_PAGE_LRU;
1654                         break;
1655                 }
1656
1657                 if (page_count(page) != 1 + page_mapcount(page)) {
1658                         result = SCAN_PAGE_COUNT;
1659                         break;
1660                 }
1661
1662                 /*
1663                  * We probably should check if the page is referenced here, but
1664                  * nobody would transfer pte_young() to PageReferenced() for us.
1665                  * And rmap walk here is just too costly...
1666                  */
1667
1668                 present++;
1669
1670                 if (need_resched()) {
1671                         cond_resched_rcu();
1672                         slot = radix_tree_iter_next(&iter);
1673                 }
1674         }
1675         rcu_read_unlock();
1676
1677         if (result == SCAN_SUCCEED) {
1678                 if (present < HPAGE_PMD_NR - khugepaged_max_ptes_none) {
1679                         result = SCAN_EXCEED_NONE_PTE;
1680                 } else {
1681                         node = khugepaged_find_target_node();
1682                         collapse_shmem(mm, mapping, start, hpage, node);
1683                 }
1684         }
1685
1686         /* TODO: tracepoints */
1687 }
1688 #else
1689 static void khugepaged_scan_shmem(struct mm_struct *mm,
1690                 struct address_space *mapping,
1691                 pgoff_t start, struct page **hpage)
1692 {
1693         BUILD_BUG();
1694 }
1695 #endif
1696
1697 static unsigned int khugepaged_scan_mm_slot(unsigned int pages,
1698                                             struct page **hpage)
1699         __releases(&khugepaged_mm_lock)
1700         __acquires(&khugepaged_mm_lock)
1701 {
1702         struct mm_slot *mm_slot;
1703         struct mm_struct *mm;
1704         struct vm_area_struct *vma;
1705         int progress = 0;
1706
1707         VM_BUG_ON(!pages);
1708         VM_BUG_ON(NR_CPUS != 1 && !spin_is_locked(&khugepaged_mm_lock));
1709
1710         if (khugepaged_scan.mm_slot)
1711                 mm_slot = khugepaged_scan.mm_slot;
1712         else {
1713                 mm_slot = list_entry(khugepaged_scan.mm_head.next,
1714                                      struct mm_slot, mm_node);
1715                 khugepaged_scan.address = 0;
1716                 khugepaged_scan.mm_slot = mm_slot;
1717         }
1718         spin_unlock(&khugepaged_mm_lock);
1719
1720         mm = mm_slot->mm;
1721         /*
1722          * Don't wait for semaphore (to avoid long wait times).  Just move to
1723          * the next mm on the list.
1724          */
1725         vma = NULL;
1726         if (unlikely(!down_read_trylock(&mm->mmap_sem)))
1727                 goto breakouterloop_mmap_sem;
1728         if (likely(!khugepaged_test_exit(mm)))
1729                 vma = find_vma(mm, khugepaged_scan.address);
1730
1731         progress++;
1732         for (; vma; vma = vma->vm_next) {
1733                 unsigned long hstart, hend;
1734
1735                 cond_resched();
1736                 if (unlikely(khugepaged_test_exit(mm))) {
1737                         progress++;
1738                         break;
1739                 }
1740                 if (!hugepage_vma_check(vma)) {
1741 skip:
1742                         progress++;
1743                         continue;
1744                 }
1745                 hstart = (vma->vm_start + ~HPAGE_PMD_MASK) & HPAGE_PMD_MASK;
1746                 hend = vma->vm_end & HPAGE_PMD_MASK;
1747                 if (hstart >= hend)
1748                         goto skip;
1749                 if (khugepaged_scan.address > hend)
1750                         goto skip;
1751                 if (khugepaged_scan.address < hstart)
1752                         khugepaged_scan.address = hstart;
1753                 VM_BUG_ON(khugepaged_scan.address & ~HPAGE_PMD_MASK);
1754
1755                 while (khugepaged_scan.address < hend) {
1756                         int ret;
1757                         cond_resched();
1758                         if (unlikely(khugepaged_test_exit(mm)))
1759                                 goto breakouterloop;
1760
1761                         VM_BUG_ON(khugepaged_scan.address < hstart ||
1762                                   khugepaged_scan.address + HPAGE_PMD_SIZE >
1763                                   hend);
1764                         if (shmem_file(vma->vm_file)) {
1765                                 struct file *file;
1766                                 pgoff_t pgoff = linear_page_index(vma,
1767                                                 khugepaged_scan.address);
1768                                 if (!shmem_huge_enabled(vma))
1769                                         goto skip;
1770                                 file = get_file(vma->vm_file);
1771                                 up_read(&mm->mmap_sem);
1772                                 ret = 1;
1773                                 khugepaged_scan_shmem(mm, file->f_mapping,
1774                                                 pgoff, hpage);
1775                                 fput(file);
1776                         } else {
1777                                 ret = khugepaged_scan_pmd(mm, vma,
1778                                                 khugepaged_scan.address,
1779                                                 hpage);
1780                         }
1781                         /* move to next address */
1782                         khugepaged_scan.address += HPAGE_PMD_SIZE;
1783                         progress += HPAGE_PMD_NR;
1784                         if (ret)
1785                                 /* we released mmap_sem so break loop */
1786                                 goto breakouterloop_mmap_sem;
1787                         if (progress >= pages)
1788                                 goto breakouterloop;
1789                 }
1790         }
1791 breakouterloop:
1792         up_read(&mm->mmap_sem); /* exit_mmap will destroy ptes after this */
1793 breakouterloop_mmap_sem:
1794
1795         spin_lock(&khugepaged_mm_lock);
1796         VM_BUG_ON(khugepaged_scan.mm_slot != mm_slot);
1797         /*
1798          * Release the current mm_slot if this mm is about to die, or
1799          * if we scanned all vmas of this mm.
1800          */
1801         if (khugepaged_test_exit(mm) || !vma) {
1802                 /*
1803                  * Make sure that if mm_users is reaching zero while
1804                  * khugepaged runs here, khugepaged_exit will find
1805                  * mm_slot not pointing to the exiting mm.
1806                  */
1807                 if (mm_slot->mm_node.next != &khugepaged_scan.mm_head) {
1808                         khugepaged_scan.mm_slot = list_entry(
1809                                 mm_slot->mm_node.next,
1810                                 struct mm_slot, mm_node);
1811                         khugepaged_scan.address = 0;
1812                 } else {
1813                         khugepaged_scan.mm_slot = NULL;
1814                         khugepaged_full_scans++;
1815                 }
1816
1817                 collect_mm_slot(mm_slot);
1818         }
1819
1820         return progress;
1821 }
1822
1823 static int khugepaged_has_work(void)
1824 {
1825         return !list_empty(&khugepaged_scan.mm_head) &&
1826                 khugepaged_enabled();
1827 }
1828
1829 static int khugepaged_wait_event(void)
1830 {
1831         return !list_empty(&khugepaged_scan.mm_head) ||
1832                 kthread_should_stop();
1833 }
1834
1835 static void khugepaged_do_scan(void)
1836 {
1837         struct page *hpage = NULL;
1838         unsigned int progress = 0, pass_through_head = 0;
1839         unsigned int pages = khugepaged_pages_to_scan;
1840         bool wait = true;
1841
1842         barrier(); /* write khugepaged_pages_to_scan to local stack */
1843
1844         while (progress < pages) {
1845                 if (!khugepaged_prealloc_page(&hpage, &wait))
1846                         break;
1847
1848                 cond_resched();
1849
1850                 if (unlikely(kthread_should_stop() || try_to_freeze()))
1851                         break;
1852
1853                 spin_lock(&khugepaged_mm_lock);
1854                 if (!khugepaged_scan.mm_slot)
1855                         pass_through_head++;
1856                 if (khugepaged_has_work() &&
1857                     pass_through_head < 2)
1858                         progress += khugepaged_scan_mm_slot(pages - progress,
1859                                                             &hpage);
1860                 else
1861                         progress = pages;
1862                 spin_unlock(&khugepaged_mm_lock);
1863         }
1864
1865         if (!IS_ERR_OR_NULL(hpage))
1866                 put_page(hpage);
1867 }
1868
1869 static bool khugepaged_should_wakeup(void)
1870 {
1871         return kthread_should_stop() ||
1872                time_after_eq(jiffies, khugepaged_sleep_expire);
1873 }
1874
1875 static void khugepaged_wait_work(void)
1876 {
1877         if (khugepaged_has_work()) {
1878                 const unsigned long scan_sleep_jiffies =
1879                         msecs_to_jiffies(khugepaged_scan_sleep_millisecs);
1880
1881                 if (!scan_sleep_jiffies)
1882                         return;
1883
1884                 khugepaged_sleep_expire = jiffies + scan_sleep_jiffies;
1885                 wait_event_freezable_timeout(khugepaged_wait,
1886                                              khugepaged_should_wakeup(),
1887                                              scan_sleep_jiffies);
1888                 return;
1889         }
1890
1891         if (khugepaged_enabled())
1892                 wait_event_freezable(khugepaged_wait, khugepaged_wait_event());
1893 }
1894
1895 static int khugepaged(void *none)
1896 {
1897         struct mm_slot *mm_slot;
1898
1899         set_freezable();
1900         set_user_nice(current, MAX_NICE);
1901
1902         while (!kthread_should_stop()) {
1903                 khugepaged_do_scan();
1904                 khugepaged_wait_work();
1905         }
1906
1907         spin_lock(&khugepaged_mm_lock);
1908         mm_slot = khugepaged_scan.mm_slot;
1909         khugepaged_scan.mm_slot = NULL;
1910         if (mm_slot)
1911                 collect_mm_slot(mm_slot);
1912         spin_unlock(&khugepaged_mm_lock);
1913         return 0;
1914 }
1915
1916 static void set_recommended_min_free_kbytes(void)
1917 {
1918         struct zone *zone;
1919         int nr_zones = 0;
1920         unsigned long recommended_min;
1921
1922         for_each_populated_zone(zone)
1923                 nr_zones++;
1924
1925         /* Ensure 2 pageblocks are free to assist fragmentation avoidance */
1926         recommended_min = pageblock_nr_pages * nr_zones * 2;
1927
1928         /*
1929          * Make sure that on average at least two pageblocks are almost free
1930          * of another type, one for a migratetype to fall back to and a
1931          * second to avoid subsequent fallbacks of other types There are 3
1932          * MIGRATE_TYPES we care about.
1933          */
1934         recommended_min += pageblock_nr_pages * nr_zones *
1935                            MIGRATE_PCPTYPES * MIGRATE_PCPTYPES;
1936
1937         /* don't ever allow to reserve more than 5% of the lowmem */
1938         recommended_min = min(recommended_min,
1939                               (unsigned long) nr_free_buffer_pages() / 20);
1940         recommended_min <<= (PAGE_SHIFT-10);
1941
1942         if (recommended_min > min_free_kbytes) {
1943                 if (user_min_free_kbytes >= 0)
1944                         pr_info("raising min_free_kbytes from %d to %lu to help transparent hugepage allocations\n",
1945                                 min_free_kbytes, recommended_min);
1946
1947                 min_free_kbytes = recommended_min;
1948         }
1949         setup_per_zone_wmarks();
1950 }
1951
1952 int start_stop_khugepaged(void)
1953 {
1954         int err = 0;
1955
1956         mutex_lock(&khugepaged_mutex);
1957         if (khugepaged_enabled()) {
1958                 if (!khugepaged_thread)
1959                         khugepaged_thread = kthread_run(khugepaged, NULL,
1960                                                         "khugepaged");
1961                 if (IS_ERR(khugepaged_thread)) {
1962                         pr_err("khugepaged: kthread_run(khugepaged) failed\n");
1963                         err = PTR_ERR(khugepaged_thread);
1964                         khugepaged_thread = NULL;
1965                         goto fail;
1966                 }
1967
1968                 if (!list_empty(&khugepaged_scan.mm_head))
1969                         wake_up_interruptible(&khugepaged_wait);
1970
1971                 set_recommended_min_free_kbytes();
1972         } else if (khugepaged_thread) {
1973                 kthread_stop(khugepaged_thread);
1974                 khugepaged_thread = NULL;
1975         }
1976 fail:
1977         mutex_unlock(&khugepaged_mutex);
1978         return err;
1979 }
1980
1981 void khugepaged_min_free_kbytes_update(void)
1982 {
1983         mutex_lock(&khugepaged_mutex);
1984         if (khugepaged_enabled() && khugepaged_thread)
1985                 set_recommended_min_free_kbytes();
1986         mutex_unlock(&khugepaged_mutex);
1987 }